复杂道路场景下智能驾驶单目3D车道线检测

打开文本图片集

关键词：自动驾驶；动态蛇形卷积；鸟瞰图；特征聚合；残差连接

中图分类号：TP391.4 文献标志码：A 文章编号：1001-3695（2025）12-036-3807-08

doi：10. 19734/j.issn.1001-3695.2025.03.0148

Monocular 3D lane detection for intelligent driving in complex road scenarios

Lou Lua†，Hu Zhenkuna，Wei Wenjiea，Shen Yia ，Wei Hanbingb （a.SchdlflVi 400074，China）

Abstract：Lanedetectionisessentialforperceptionanddecision-makinginintellgentvehicles.Toadressthechallngehat existingmethodsstruggle tobalance detectionaccuracyandspeed incomplex traficscenarios，this paper proposedanfficient monocular 3D lane detection method.The proposed method leverageddynamicdeformable convolution toefectivelycapture the curvedand elongated shapefeaturesof lanes.It employed multi-scaleviewtransformationsand feature aggregatio toreducefeaturelossandobtainlaneposition informationatdiferentlevels.Furthermore，itintroducedanauxiliarysupervision strategywithesidualconections toenancethemodel’srepresentationcapabilityExperimentalresultson3Dlanedetection benchmarks show that the proposed method achieves an F1 -score of 98.6% on the synthetic Apollo 3D dataset，outperforming BEV-LaneDet and LATR by1. 7% and 1. 8% ，respectively.On the large-scale real-world OpenLane dataset，the method reaches an F1 -score of 59.8% ，which is 1.4% higher than BEV-LaneDet. The performance gains are particularly notable in uphill/downhill and curved road scenes，with improvements of 4. 6% and 2. 5% ，respectively.Although the F1 -scoreis （20 2.1% lowerthan the Transformer-basedLATRmethod，the methodruns at 8O.1fps，which is 5.7times fasterthanLATR’s 14 fps.These results demonstrate thatthe proposed method can improve theaccuracyand robustnessof 3D lane detection in complex scenes，and achieve a favorable balance between detection performance and running inference speed.

Key words：autonomous driving；dynamicsnakeconvolution；bird’s-eye-view；feature agregation；residualconection

0 引言

车道线是一种路面主要交通标志，具有划分道路、引导行驶、保证安全等功能。（剩余20467字）